Does the Stock Market Fully Appreciate the Implications of Leading Indicators for Future Earnings? Evidence from Order Backlog

Transcription

1 Does the Stock Market Fully Appreciate the Implications of Leading Indicators for Future Earnings? Evidence from Order Backlog Shivaram Rajgopal Department of Accounting University of Washington Box Seattle, WA Fax: Terry Shevlin* Deloitte and Touche Professor Department of Accounting University of Washington Box Seattle, WA Tel: Fax: and Mohan Venkatachalam Fuqua School of Business Duke University P.O. Box Durham, NC Tel: Fax: June 11, 2003 *Contact author

2 Abstract. Although leading indicators are becoming increasingly important for equity valuation, disclosures of such indicators suffer from the absence of GAAP related guidance on content and presentation. We explicitly examine (i) whether one leading indicator order backlog predicts future earnings, and (ii) whether market participants correctly incorporate such predictive ability in determining share prices. We find that the stock market overweights the contribution of order backlog in predicting future earnings, and a hedge strategy that exploits such overweighting generates significant future abnormal returns. However, such mispricing is not due to analysts inability to incorporate order backlog into their earnings forecasts. Keywords: leading indicator, backlog, mispricing, market efficiency, analyst forecasts JEL classification: M41, G14

3 In this study, we investigate whether capital market participants, i.e., investors and analysts, incorporate the implications of one leading indicator order backlog for future earnings in an efficient manner when determining stock prices and earnings forecasts. The market value of a firm s shares and analyst forecasts should reflect an unbiased expectation of future earnings. To the extent current earnings persist in the future, the link between current and future earnings is relatively apparent. However, in today s dynamic business environment, a substantial portion of future earnings may reflect the contribution of several leading indicators such as backlog (Behn, 1996), product market share (Amir and Lev, 1996), customer satisfaction (Ittner and Larcker, 1998a), patents (Deng, Lev, and Narin, 1999), web traffic (Trueman, Wong, and Zhang, 2001; Rajgopal, Venkatachalam, and Kotha, 2003) or managerial actions (Rajgopal, Venkatachalam, and Kotha, 2002). Under generally accepted U.S. accounting principles, such leading indicators are not recognized but are often disclosed in the firm s financial statements. When an economically significant portion of a firm s future earnings stems from leading indicators, the absence of explicit information about the contribution of such indicators to future earnings complicates the task of equity valuation. Disclosure issues related to leading indicators have grown in importance in recent years. One reason for such focus is claims about the declining importance of accounting-based performance metrics such as earnings and equity book values (see Collins, Maydew, and Weiss, 1997; Francis and Schipper, 1999; Brown, Lo, and Lys, 1999; Lev and Zarowin, 1999). Further, the Jenkins Committee report (AICPA, 1994) and research on the lead-lag relation between stock prices and earnings (see Kothari, 2001) kindled an active interest among academics in understanding whether leading indicators of future earnings could potentially remedy the perceived fall in the value-relevance of earnings and book value. Several practitioners (e.g., 1

4 Kaplan and Norton, 1996; Eccles et al., 2001) have called for a reporting framework that integrates traditional financial measures and leading indicators. Recently, national regulators such as the FASB (2001) and SEC (2001) and international regulators such as the Canadian Institute of Chartered Accountants (1995) and Institute of Chartered Accountants of England and Wales (2000) have published studies that discuss the possibility of mandating enhanced disclosures of leading indicators. The increased importance of non-gaap leading indicators raises the question of whether security prices reflect their unbiased contribution to future earnings. In an efficient market, the stock market correctly impounds the value of a leading indicator for future earnings. Hence, there ought to be no association between leading indictors and future stock returns. On the other hand, several characteristics of these leading indicators impede a clear understanding of their implications for future earnings. Most leading indicators are inherently idiosyncratic to specific industries and some are even specific to individual firms. Disclosure of such metrics is voluntary and is open to management discretion. Furthermore, the content and presentation of such disclosures are not standardized by GAAP. Many indicators are not denominated in dollars, rendering cross-firm comparison of such indicators difficult. Moreover, a recent FASB study (FASB 2001) found several problems in the extant disclosures: (i) indicators were presented in isolation without relating them to the context of the industry and the company; (ii) the measurement and presentation of some indicators were not consistent over time; (iii) some firms presented vague metrics such as organization enhancing customers or image enhancing customers without a concise definition of the metric or its measurement. These complications raise the possibility that stock prices do not fully incorporate the contribution of leading indicators for future earnings. On account of uncertainty inherent in 2

5 translating leading indicators to value creation, investors may excessively discount the contribution of such indicators to future earnings. Underpricing may also arise if investors are functionally fixated on current earnings -- i.e., they mechanically accept firms current earnings numbers at face value without adjusting for the long-term benefits of the underlying activities or processes captured by leading indicators. In contrast, some observers have suggested that investors overestimate the benefits from leading indicators, especially in hot sectors of the economy such as Internet-intensive businesses or the wireless industry (Damodaran, 2001; D Avolio, Gildor, and Shleifer, 2001). There is wide coverage of leading indicators such as POPS (population coverage), book-to-bill ratio or customer satisfaction indices in the media and some firms devote intensive marketing efforts to promote such indicators. As a result, the market may be overly optimistic about the contribution of such indicators to future earnings. The objective of our study is to explicitly examine whether investors appropriately account for the contribution of leading indicators to future earnings. We choose order backlog as the leading indicator in our empirical tests because an investigation into the potential mispricing of a widely disseminated, dollar denominated leading indicator that is readily available and cross-sectionally comparable for a large number of firms across several industries gives the null of market efficiency the best possible chance of success. Moreover, order backlog is economically significant (30% of total assets for the median sample firm) and is somewhat standardized in its definition and disclosure. If the market has difficulty in interpreting the implications of order backlog for future earnings, it is more than likely that the market fails to fully appreciate the implications of non-dollar denominated, less widely disseminated, and less objective leading indicators such as customer satisfaction measures, web 3

6 traffic, patents or firm-or industry-specific metrics such as manufacturing cycle time or time-tomarket measures. Our empirical tests use the Mishkin (1983) framework to examine whether market participants incorporate into stock prices the predictive ability of order backlog for future earnings. The results indicate that the stock market behaves as if backlog contributes more to future earnings than is implied by the cross-sectional association between backlog and future earnings. To corroborate findings from the Mishkin (1983) test, we assess whether abnormal future returns can be earned by taking positions on current order backlog. We find that a Fama and MacBeth (1973) type hedge strategy based on taking positions on order backlog can earn abnormal returns of 7.7% as compared to 16.6% abnormal returns that can be earned from Sloan s (1996) accrual-based anomaly. Moreover, abnormal returns based on order backlog are robust to the inclusion of Fama and French (1992) risk factors and are incremental to abnormal returns earned from Sloan s (1996) accrual anomaly. Next, we examine whether the failure of equity investors to correctly incorporate the implications of order backlog for future earnings is at least partially explained by the failure of sophisticated information intermediaries, such as equity analysts, to incorporate the contribution of backlog to future earnings when generating earnings forecasts. Our results suggest otherwise; we find that median consensus analysts forecasts correctly incorporate the information contained in backlog for future earnings. Finally, we find that even after controlling for information in analysts forecasts for future earnings, investors continue to place weight on order backlog information. This result is consistent with market participants not appreciating the fact that analyst forecasts already incorporate information in order backlog. 4

7 Our study extends two areas of prior research. First, we contribute to the literature that documents a relation between stock prices and leading indicators. The association between leading indicators and stock prices is usually taken as evidence that the market correctly prices the contribution of such leading indicators to future earnings. However, recent research raises doubts about that assumption. In particular, survey evidence presented by Ittner and Larcker (1998b) suggests that even managers question the quality of leading indicator information. Furthermore, a field study conducted by Banker, Potter, and Srinivasan (2000) indicates that managers, let alone investors, cannot fully articulate the link between leading indicators and firm value. The evidence we present suggests that the market misprices leading indicators although sophisticated information intermediaries such as analysts fully appreciate these disclosures. To the extent such abnormal returns imply that there are costs associated with the limited disclosure of information on order backlog (such as estimated future earnings from backlog), we believe that our findings can potentially inform standard setting bodies during their deliberations on standardizing or enhancing the disclosure of non-gaap leading indicators (see FASB 2001, SEC 2001). Second, we extend prior research by Sloan (1996) that examines market efficiency with respect to earnings, accruals, and cash flows, to leading indicators and their contribution to future earnings. The remainder of the paper is organized as follows. In Section 1, we briefly describe the difficulties associated with disclosures of leading indicators, prior research and institutional details on order backlog. Section 2 discusses the sample while section 3 presents the experimental procedures employed and the findings. Section 4 investigates whether equity analysts appreciate the role of backlog in forecasting future earnings and how the stock market 5

8 interprets backlog in the presence of analyst forecasts. Section 5 provides some concluding remarks. 1. Background 1.1. Disclosure Issues on Leading Indicators In recent years, academics, practitioners and regulatory bodies have called for increased disclosure of leading indicators by firms (e.g., AICPA, 1994; Eccles et al., 2001; Lev, 2001). These groups argue that traditional financial measures such as book value and earnings have reduced relevance for valuation purposes, especially in rapidly evolving technology-based industries. 1 Survey evidence suggests that professional analysts use leading indicators reported in firms financial reports (Previts et al., 1994). Concurrently, several firms have adopted the use of leading indicators for internal performance evaluation (Ittner and Larcker, 1998b). Given the increasing importance of leading indicators for valuation and managerial compensation purposes, many practitioners have called for a disclosure framework that integrates leading indicators with traditional financial measures (Kaplan and Norton, 1996; Eccles et al., 2001). Extant disclosures of leading indicators are subject to several limitations (Eccles et al., 2001; FASB, 2001). First, disclosures of leading indicators are voluntary in nature and are hence not available for all firms. Second, the types of measures disclosed and the reporting formats used vary both in the cross-section and across time. Third, leading indicators are often industryspecific and sometimes even firm-specific. Fourth, survey evidence reported in Ittner and Larcker (1998b) indicates that (i) executives have concerns about the quality of leading indicators in certain areas such as employee performance, environment and innovation; and (ii) executives view financial information to be of higher quality than leading indicator data. Finally, two recent studies cast doubt on whether managers, let alone investors, can fully articulate the 6

9 links between leading indicators and firm value (Banker, Potter, and Srinivasan, 2000; Ittner and Larcker, 1998b). For example, in a field study covering the hotel industry, Banker, Potter, and Srinivasan (2000) find that while managers are aware of a relation between customer satisfaction and firm profitability, they did not fully appreciate the nature and the timing of this relation. These limitations raise questions about the relevance and reliability of disclosures related to leading indicators. Research on the value-relevance of leading indicators suggests that such indicators are perhaps relevant and reliable enough for researchers to detect an association between leading indicators and contemporaneous stock prices. In the wireless industry, Amir and Lev (1996) find that income statement and balance sheet information explain little of the cross-sectional variation in the market values of wireless firms. However, two leading indicators service area population (a product market size measure) and population penetration (a product market share measure) are highly value-relevant incremental to traditional financial measures. For a broad cross-section of industries, Behn (1996) finds that unfulfilled order backlogs are value-relevant. Other examples of research that documents value-relevance of leading indicators include Ittner and Larcker (1998a) on customer satisfaction, Chandra, Procassani, and Waymire (1999) on the book-to-bill ratio in the semi-conductor industry, Deng, Lev, and Narin (1999) on patent counts, Hughes (2000) on sulfur dioxide emissions, and Francis, Schipper and Vincent (2003) on airline load factors, home building order backlog, and same stores sales. Research on Internet firms related to web traffic, loyalty, and the quality of online customer experience (e.g., Trueman, Wong, and Zhang, 2001; Demers and Lev, 2001; Rajgopal, Venkatachalam, and Kotha, 2002, 2003) highlights the importance of leading indicators for emerging high-technology firms. 7

10 Most of the cited value-relevance research rests, however, on two important assumptions. First, the value-relevance of the leading indicator is usually assumed to be indicative of future earnings or future growth options in the firm. Many of the cited studies do not test whether these leading indicators predict realized future earnings. 2 Second, all the cited studies assume market efficiency in assessing the relevance and reliability of leading indicators. However, the assumption of market efficiency, even with respect to widely disseminated financial measures such as earnings and accruals, has been challenged (e.g., Bernard and Thomas, 1990; Sloan, 1996). Further, a recent commentary by the AAA Financial Accounting Standards Committee (2002, p. 357) raises questions about investors ability to use leading indicators appropriately in firm valuation. We explicitly examine both these assumptions in the current study. In particular, we model the association between a leading indicator (in our case, order backlog) and one-yearahead realized earnings and assess whether stock prices fully reflect this association. We also assess whether analyst forecasts incorporate the contribution of backlog to future earnings and whether the stock market correctly interprets such analyst forecasts. To the extent that market prices do not appropriately reflect the implications of such leading indicators for future earnings, our results suggest that information in leading indicators can be exploited to earn abnormal returns Institutional Details about Order Backlog Order backlog represents contractual orders that are unfulfilled but are scheduled to be executed in later accounting periods. Item 101(c) (VIII) of SEC regulations S-K requires disclosure, to the extent material, of the dollar amount of the order backlog believed to be firm, as of a recent date and as of a comparable date in the preceding fiscal year, together with an indication of the portion 8

11 thereof not reasonably expected to be filled within the current fiscal year, and seasonal or other material aspects of backlog. Order backlog information is usually neither audited nor reviewed, although under AICPA Statement on Auditing Standards No. 8 the auditor is required to read the SEC 10-K filing to determine whether it is consistent with the financial statements. Order backlog disclosures are commonly concentrated in a few industries such as durable manufacturing and computers (see sample description in section 2). These disclosures usually include data about backlog at the end of the current and previous fiscal year, the percentage likely to be recognized in revenues in the subsequent year and backlog numbers for significant segments of the business. To illustrate, Motorola s 10-K for the fiscal year ended December 31, 2000 contains the following disclosure about its backlog position: Motorola's aggregate backlog position, including the backlog position of subsidiaries through which some of its business units operate, as of the end of the last two fiscal years, was approximately as follows: December 31, $9.62 billion December 31, $9.92 billion Except as previously discussed in this Item 1, the orders supporting the 2000 backlog amounts shown in the foregoing table are believed to be generally firm, and approximately 94% of orders on hand at December 31, 2000 are expected to be shipped during However, this is a forward-looking estimate of the amount expected to be shipped, and future events may cause the percentage actually shipped to change. Motorola s backlog equals 26% of its fiscal 2000 sales and 23% of total assets at 2000 fiscal year-end. It is also interesting to note that Motorola expects 94% of its backlog to be reflected in next year s earnings. In contrast, Lockheed Martin s backlog will take longer to get reflected in future earnings, as is evident from the following disclosures made by the company in its fiscal K: At December 31, 2000, our total negotiated backlog was $56.4 billion compared with $45.9 billion at the end of 1999 Of our total 2000 year-end backlog, approximately $40.7 billion, or 72%, is not expected to be filled within one year. These amounts include both unfilled firm orders for our products for which funding has been both authorized and appropriated by the customer (Congress in the case of U.S. Government agencies) and firm orders for which funding has not been appropriated. 9

12 Lockheed Martin s backlog for fiscal 2000 is 222% of sales and 186% of total assets. Two observations can be made from a review of these disclosures. First, backlog numbers are economically significant enough to expect market participants to pay attention to them when forecasting a firm s future earnings. Second, the number of years of annual earnings that would be affected by backlog at any year-end is likely to vary by industry. Our empirical tests assume that, in the cross-section, year-end backlog substantially feeds into one-period ahead future earnings. We report results related to one lag because only a small proportion (7%) of our sample (discussed in the next section) have a backlog-to-annual sales ratio in excess of 100%. Although firms are required by Item 101(c) (VIII) of SEC regulations S-K to disclose the proportion of backlog not scheduled to be executed in the next fiscal year, Compustat does not code that information. Hence, we are unable to use that data in our empirical tests. However, unreported analyses suggest that backlog follows a random-walk process. Thus, the assumption that backlog maps into future earnings with just one lag seems reasonable. We relax the one-lag structure assumption in our sensitivity analyses. Even if backlog is economically significant, market participants may not use backlog disclosures unless such disclosures are reliable. Previous studies provide evidence that backlog disclosures are value-relevant in general (Behn, 1996; Chandra, Procassini, and Waymire, 1999), and are useful in predicting future sales (Liu, Livnat, and Ryan, 1996). Lev and Thiagarajan (1993) use the proportion of sales changes to order backlog changes as a revenue management proxy to predict (and find) a negative contemporaneous association between returns and such proportion. However, none of the cited papers tests whether market participants fully appreciate the association between order backlog and future earnings. 2. Sample Selection, Variable Measurement and Descriptive Statistics 10

13 In this section, we provide descriptive data for the sample, which includes all firm-years satisfying our data requirements. The empirical analyses use firms with non-zero order backlog information during the period available from the 2000 Compustat active and research data tapes. We obtain 29,574 firm-year observations for which financial statement data are available. We eliminate 7,218 firm-year observations for which stock price and return data are unavailable from the 2000 CRSP tapes. To avoid the undue influence of extreme observations we eliminate 465 firm-years in the extreme upper and lower 1% of the order backlog distribution. Thus, our final sample comprises 21,891 firm-year observations representing 3,170 firms. For analysis that requires accruals and operating cash flows data we use a reduced sample of 13,012 observations because we are able to obtain operating cash flows from SFAS 95 disclosures only since Primary Variables BKLG t = Backlog (Compustat #98) in year t scaled by average total assets (calculated as the arithmetic mean of total assets at the beginning and end of the fiscal year). EARN t = Income before extraordinary items (Compustat #18) in year t scaled by average total assets. OPCF t = Operating cash flows (Compustat #308) in year t scaled by average total assets. ACC t = Accruals measured as the difference between EARN t and OPCF t. SAR t+1 = Size-adjusted security return, measured as the realized security return from CRSP, cumulated over 12 months beginning with the fourth month after the end of fiscal year t, minus the corresponding mean return for all CRSP firms in the same market capitalization decile at the end of year t. 11

14 Following Sloan (1996), we control for Fama and French (1992) factors (size, book-tomarket and beta) and prior-year earnings scaled by share price (Basu, 1977) to rule out alternate explanations for any anomalous security returns. In particular, we consider four variables: SIZE t = the natural log of the market-value of common equity (MVE), measured at four months after fiscal year-end. LnBM t = the natural log of the ratio of book value of common equity to the market value of common equity (BM), measured at four months after fiscal year-end. 3 BETA t = the CAPM beta estimated from a regression of raw monthly returns on the CRSP value-weighted monthly return index over a period of 60-months ending four months after each firm s fiscal year-end. 4 EP t = earnings per share divided by stock price measured at four months after fiscal year-end. Descriptive data related to these variables are reported in Panels A and B of Table 1. Backlog for the average (median) firm is 46% (30%) of average total assets. The earnings for an average (median) firm are 0% (4%) of average total assets. Backlog is a large proportion of average total assets relative to earnings (as a proportion of total assets) because backlog represents future sales. On average, accruals are negative, consistent with depreciation being a dominant accrual for these firms. Panel C presents data on the potential relation between order backlog deflated by assets (i.e., BKLG) and other firm characteristics such as backlog deflated by sales, earnings, cash flows, and accruals, Fama and French (1992) risk factors, and other variables shown to predict future returns. In particular, we form ten portfolios sorted on order backlog (deflated by assets) and report the means of various characteristics of firms in these portfolios. We find that backlog deflated by sales closely follows the distribution of backlog deflated by assets. Other than that, there are no clear patterns between BKLG and the reported 12

15 firm characteristics. Thus, the descriptive evidence presented in panel C indicates prima facie that BKLG is not strongly related to other well documented anomalies, although we investigate this issue in greater depth later in the paper. The correlations reported in panel D confirm that indication. In Panel E we provide the industry composition for our sample firms. Note that a substantial number of observations (over 70% of the sample) come from two industries - durable manufacturing and computers (discussed further below). ----Table 1 about here Empirical Analyses We present the empirical tests in several stages. In sections 3.1 and 3.2 we estimate the historical relation between backlog and future earnings, and the relation between backlog and future earnings implicit in security prices. A comparison of these historical and market-inferred weights using the Mishkin (1983) test indicates whether investors correctly appreciate the importance of order backlog for future earnings. In section 3.3 we investigate whether abnormal returns can be earned by exploiting investors misweighting of the contribution of order backlog to future earnings. In section 4, we examine whether equity analysts use backlog information in an efficient manner while forecasting future earnings. Finally, we also assess whether market inefficiencies with respect to backlog are present after controlling for information in analysts forecasts Mishkin Test Our initial empirical tests employ the Mishkin (1983) framework to test whether the stock market is efficient in impounding the information contained in backlog for future earnings. This framework, introduced by Sloan (1996) to the accounting literature, has since been used by a number of studies that test for market efficiency. We infer mispricing if the valuation coefficient 13

16 attributed to backlog by market participants is different from the weight in the association between backlog and future earnings. Similar to prior research, we jointly estimate a forecasting specification for the leading indicator and the rational expectations pricing specification. We extend the earnings forecasting equation in Sloan (1996) to incorporate the implications of order backlog for future earnings as follows (firm subscripts suppressed): EARN t+1 = ω 0 + ω 1 EARN t + ω 2 BKLG t + ε t+1 (1) The coefficient ω 1 represents earnings persistence while ω 2 captures the incremental contribution of current order backlog for future earnings. Note that backlog is denominated in sales dollars. Hence, ω 2 incorporates the joint effect of i) the probability of conversion of order backlog into future sales; and ii) the accounting rate of return (margins) on such converted sales. For the purposes of our analyses, we assume ω 2 is a cross-sectional constant. 5 Next, we assume that the market reacts to unexpected earnings conditioned on last year s earnings and order backlog. That is: SAR t+1 = β 0 + β 1 UEARN t+1 + ν t+1 (2) where UEARN t+1, the unexpected earnings variable, is decomposed into realized earnings, EARN t+1, and expected earnings, E(EARN t+1 ) with E(.) as the expectation operator: UEARN t+1 = EARN t+1 E (EARN t+1 ) (3) We substitute the linear prediction of EARN t+1 in equation (1) for expected earnings E(EARN t+1 ) embedded in equation (3), and rewrite equation (2) as: SAR t+1 = β 0 + β 1 (EARN t+1 ω 0 * - ω 1 * EARN t - ω 2 * BKLG t ) + ν t+1 (4) We refer to equation (4) as the returns equation. Market efficiency with respect to backlog imposes the constraint that ω 2 * from the returns equation (4) is the same as ω 2 from 14

17 forecasting equation (1). This nonlinear constraint requires that the stock market rationally anticipate the implications of current order backlog for future earnings. If ω 2 > ω 2 *, the market assesses a lower contribution of backlog to future earnings than that warranted by the underlying cross-sectional association of backlog and future earnings. On the other hand, if ω 2 < ω 2 *, the market assesses a higher contribution of backlog to future earnings than is warranted. The two equations in the system are estimated using iterative weighted non-linear least squares (Mishkin 1983). Market efficiency is tested using a likelihood ratio statistic which is distributed asymptotically chi-square (q): where 2*n* ln (SSR c / SSR u ) (5) q = the number of constraints imposed by market efficiency, n = the number of observations in each equation, SSR c = the sum of squared residuals from the constrained weighted system, and SSR u = the sum of squared residuals from the unconstrained weighted system Mishkin Test Results Our initial analyses compare the historical relation between backlog and realized earnings to the relation between backlog and earnings expectations embedded in security prices. Table 2 reports the historical relation between realized earnings, EARN t+1, and order backlog, BKLG t, and the stock market s implicit weighting of order backlog. The coefficient on earnings in the forecasting equation, ω 1 is and is statistically different from zero. This coefficient represents the persistence of the accounting rate of return on assets and is less than unity, implying thereby that accounting rates of return are mean-reverting (Beaver, 1970; Freeman, Ohlson, and Penman, 1982). The coefficient on earnings in the returns equation (4), ω 1 * is 0.536, but is only weakly different from its counterpart in the forecasting equation (1): the likelihood 15

18 test for market efficiency on annual earnings (ω 1 = ω * 1 ) is a marginal 3.81 (significance level = 0.05). Thus, we detect traces of inefficiency in the market s evaluation of earnings Table 2 about here --- Turning to the role of backlog, we find that the coefficient on BKLG in the forecasting equation, ω 2, is positive (coefficient = 0.008) and statistically significant (p < 0.01). While the magnitude of the coefficient on order backlog (0.008) may appear to be small relative to the magnitude on the earnings coefficient (0.536), recall that order backlog is denominated in sales dollars. To convert order backlog into comparable earnings dollars, we divide the coefficient on backlog, 0.008, by the median return on sales for our sample firms (2.6%). This conversion is equivalent to estimating the backlog coefficient after multiplying the backlog variable by the median return on sales. The resulting quotient is a coefficient of Thus, the incremental impact of order backlog for future earnings (0.31) is comparable in magnitude to the coefficient on lagged earnings (0.536). The impact of backlog is lower, probably because some of the backlog may not convert to future sales or may convert at lower margins. This suggests that order backlog is incrementally informative about future earnings even after controlling for information in past earnings. Of most importance is that the coefficient on BKLG in the returns equation, ω 2 *, is over four times larger than that in the forecasting equation and is statistically significant (coefficient = 0.039, p < 0.01). The difference between the two coefficients is statistically significant (likelihood ratio test statistic = 12.06, p < 0.00). Thus the stock market appears to place a higher weight on order backlog relative to the weight inherent in the association between backlog and future earnings Abnormal Returns to Hedge Strategy 16

19 The preceding section uses the Mishkin test to demonstrate that the stock market overprices backlog information. However, the Mishkin test is not direct evidence of market inefficiency for at least two reasons (see Beaver and McNichols, 2001; Wahlen, 2001). First, model misspecification or the assumptions implicit in the Mishkin procedure (such as linearity of the model specification, and efficient stock market pricing of omitted variables) may cause an illusion of market inefficiency (see Kraft, Leone, and Wasley, 2001). Second, because the coefficients are estimated from a set of contemporaneous observations, the procedure suffers from a look-ahead bias. Investors do not know the implications (i.e., the weight) of backlog for future earnings until later in the sample period. Hence, we turn to an additional test, the ability of portfolios formed on the cross-sectional distribution of backlog to predict future abnormal returns. The strategy we implement relies on the construction of zero-investment portfolios (Fama and MacBeth, 1973). Portfolios are formed as follows: First, for each year from 1981 to 1999, we calculate the scaled decile rank for BKLG t for each firm. In particular, we rank the values of BKLG t into deciles (0,9) each year and divide the decile number by nine so that each observation related to BKLG t takes a value ranging between zero and one. We estimate the following crosssectional OLS regression for each of the 19 years in the sample: SAR t+1 = γ 0 + γ 1 BKLG dec t + φ t+1 (6) The basic idea behind the Fama and MacBeth (1973) regressions is to project the sizeadjusted returns on the intercept and the BKLG dec t variable for each cross section and then aggregate the estimates over the 19 years. Tests of statistical significance of γ 1 are based on the standard error calculated from the distribution of the individual yearly coefficients. This test 17

20 overcomes bias due to cross-sectional correlation in error terms but assumes independence in error terms across time (Bernard, 1987). Coefficient γ 1 represents the size-adjusted abnormal return to a zero-investment portfolio optimally formed to exploit the information in the backlog variable. This is because the weights assigned to each firm in the BKLG dec variable, represented by the rows of the matrix (X'X) -1 X', where X = [1, BKLG dec t], sum to zero. Because the portfolio weights are determined without foreknowledge of future abnormal returns, we are investigating executable trading strategies. Our strategy involves taking positions in firms that have reported results within 4 months of the fiscal-year end to allow for the determination of portfolio weights from (X'X) -1 X' used to ascertain the investment positions. Thus, firms receiving negative weights are sold short and firms with positive weights are bought. The long and short positions are closed after one year. The abnormal returns, represented by coefficient γ 1, are comparable to abnormal returns to a zero-investment portfolio with long (short) positions in firms within the lowest (highest) deciles of backlog (see Bernard and Thomas, 1990). Results from the Fama-MacBeth regressions reported in Table 3 generally confirm the findings from the Mishkin test. There is a negative relation (coefficient = ) between BKLG and future returns that is statistically significant (t-statistic of 2.38, p < 0.05). The negative sign on the coefficient on BKLG is consistent with the difference in historical and security-market weightings of the contribution of current backlog to future earnings documented using the Mishkin framework. Because the market overestimates the future earnings implications of backlog we should observe negative abnormal returns for portfolios ranked on backlog. Thus, the abnormal return of 5.8% to a trading strategy based on order backlog is both statistically and economically significant. 18

21 --- Table 3 about here Robustness Checks and Sensitivity Analyses Controlling for Fama-French Factors and the Basu (1977) Anomaly Prior research has shown that future abnormal returns are associated with other variables, such as firm size (SIZE), book-to-market ratio (BM), and systematic risk (BETA). Fama and French (1992) conjecture that these variables reflect unknown risk factors and, hence, are associated with future expected returns. Although panels C and D of Table 1 report the absence of strong relations between BKLG and the Fama-French factors, to be consistent with standard practice in recent asset-pricing literature, we investigate whether abnormal returns related to backlog are independent of returns observed in connection with the Fama-French factors. We also include the earnings-to-price ratio (EP) to control for the earnings-price anomaly documented by Basu (1977). To assess whether the backlog anomaly generates incremental returns to the Fama-French factors and the Basu anomaly, we estimate the following regression: SAR t+1 = γ 0 + γ 1 BKLG dec t + γ 2 SIZE dec t + γ 3 BETA dec t + γ 4 LnBM dec t + γ 5 EP dec t + ϕ t+1 (7) where SIZE dec t, BETA dec t, and LnBM dec t relate to scaled decile ranks (ranging from 0 to 1) for the three Fama-French factors. EP dec t refers to scaled decile ranks of the earnings-price ratio. In the regression specification (7), coefficient γ 1 represents the incremental size-adjusted abnormal return to a zero-investment portfolio in the backlog variable. Results reported in panel B of Table 3 indicate that incremental abnormal returns related to BKLG persist after controlling for potentially confounding variables. However, controlling for Fama-French factors and the Basu anomaly reduces the incremental return to the BKLG based hedge portfolio to 4.8% from 5.8% in panel A with no controls. Turning to the control 19

22 variables, it is interesting to note that there is a statistically significant size-adjusted incremental return of 8.3% to the SIZE based hedge portfolio. Although this result is somewhat curious, it is consistent with Foster, Olsen, and Shevlin (1984), Bernard (1987), and Shevlin and Shores (1993) who find that the use of size-adjusted raw returns does not appear to ensure a perfect control for size. We also find a positive incremental return to the LnBM portfolio of 7.5%. The signs of abnormal returns on the SIZE and LnBM portfolios are consistent with those observed by Fama and French (1992) Controlling for the Accrual Anomaly In this section, we assess whether the BKLG related anomaly is distinct from the anomaly documented by Sloan (1996). Sloan shows that a hedge portfolio strategy based on accruals can earn abnormal returns in subsequent periods because the stock market overweights the persistence of the accruals for future earnings. Sloan (1996) and a few subsequent papers such as Xie (2001) infer the accruals component of earnings from balance sheet accounts such as current assets and current liabilities. However, Hribar and Collins (2002) demonstrate that accruals calculated from cash flow numbers reported by firms under SFAS 95 are less susceptible to the contaminating influences of acquisitions, mergers, and divestitures than accruals deduced from balance sheet information. Hence, we further restrict our sample to only firm-years for which SFAS 95 cash flow data are available. In Table 4, we repeat the analysis of the forecasting equation (1) and returns equation (4) on the reduced sample of 13,012 observations after decomposing earnings into the SFAS 95 derived cash flow component scaled by average assets (OPCF) and accruals scaled by average assets (ACC). Similar to Sloan (1996) we find that the market overweights the accruals component and underweights the cash flow component of total earnings in the returns equation. 20

23 Specifically, the coefficient on operating cash flows in the forecasting equation, ω 1a, is 0.726, whereas the coefficient on operating cash flows in the returns equation, ω 1a *, is and the difference is significant at conventional levels (likelihood ratio statistic = 11.84, p < 0.01). The coefficient on accruals in the forecasting equation, ω 1b, is whereas its counterpart in the returns equation, ω 1b *, is (likelihood ratio statistic = 15.31, p < 0.01). Most important, however, the market continues to overweight BKLG in the returns equation (ω 2 * is 0.052) compared to the weight on BKLG in the forecasting equation (ω 2 is 0.005) and the difference is statistically significant (likelihood ratio statistic is 13.91, p < 0.00). Thus, the BKLG anomaly is incremental to the accruals anomaly documented by Sloan (1996). --- Table 4 about here --- To ensure that abnormal returns can be earned from the BKLG anomaly incremental to those from the accrual anomaly, we augment the Fama-Macbeth type equation (7) with an additional regressor, ACC dec t. This variable represents accruals for year t divided by average assets ranked into deciles (0,9) each year and scaled by nine so that each observation takes a value ranging between zero and one. Panel A of Table 5 reports the results of assuming hedge positions on order backlog for the reduced sample of firm-years with SFAS 95 data. It turns out that the hedge portfolio return for the reduced sample is higher at 8.1% as compared to the full sample return of 5.8% reported earlier. Systematic differences between the reduced sample and the full sample could potentially account for the higher return on the backlog variable in the reduced sample. For example, we find that firms in the reduced sample are systematically larger on average (t-statistic to assess difference in means is 11.43) and have smaller CAPM betas (t-statistic for difference of means is 7.54). 21

24 --- Table 5 about here --- Panel B of Table 5 shows that the introduction of Fama-French factors and both the Sloan and the Basu anomaly do not affect abnormal returns from the BKLG anomaly. 9 A zeroinvestment strategy on BKLG variable earns incremental abnormal returns of 7.7%. 10 In contrast, Sloan s accrual strategy appears to earn an incremental abnormal return of 16.6%. Notice that the negative coefficient on the accruals variable is consistent with the market overreacting to accruals. Thus, the BKLG strategy earns as much as 46% of the returns from Sloan s anomaly Additional Sensitivity Analyses In this section we conduct several additional tests to further assess the robustness of our results. We begin by considering the sensitivity of our analysis to the assumption of a single-lag relation of backlog for future earnings. It is quite likely that backlog follows a more complicated process and investors may also use a more sophisticated expectations model; as a result, our estimates of the weights implicit on backlog in the estimation equation (ω 2 ) and valuation weight placed by investors on backlog (ω 2 *) may be biased. However, following Mishkin (1983) and Abel and Mishkin (1983), Sloan (1996) argues that the tests of cross-system non-linear constraints remain valid tests of market efficiency regardless of whether the forecasting equation has omitted variables (such as a second-lag backlog term). Abel and Mishkin (1983) and Mishkin (1983) also point out that if the model generating the dependent variables (earnings at t+1 in our setting) is incorrectly specified, then the standard errors-in-variables problem arises, meaning estimates of the valuation equation weights, ω 1 * and ω 2 *, will be inconsistent, and the test will have lower power. Rejection of the null hypothesis of efficiency with a test of known low power would be an even stronger case against efficiency. Thus, our inferences about the degree to which 22

25 information in prior backlogs is incorporated into the earnings expectations process are likely to be robust to the actual time series process and the form of investor expectations. Furthermore, as Soffer and Lys (1999) point out in their appendix 3, ω 2 can be viewed as a weighted average of the extent to which future earnings are related to prior backlog lags that are correlated with the first lag. Similarly, ω 2 * can be viewed as a weighted average of the extent to which investors have incorporated information from prior backlog lags that are correlated with the first lag. Thus we capture, in a single measure, the degree to which all information correlated with the first lag is incorporated into earnings expectations. Considering that the correlation between first-lag of backlog and the second-lag is very high (ρ = 0.87) in our sample, we argue that modeling the first lag alone is reasonable for the purpose of assessing deviation from market efficiency with respect to the information in order backlog. Notwithstanding the above arguments, we incorporate a second-lag of backlog into the earnings estimation equation (1) and (4) and test whether the weights placed by the market on both lags of backlog in equation (4 ) below are consistent with weights in the forecasting equation (1 ). EARN t+1 = ω 0 + ω 1 EARN t + ω 2 BKLG t + ω 3 BKLG t-1 + ε t+1 (1 ) SAR t+1 = β 0 + β 1 (EARN t+1 ω 0 * - ω 1 * EARN t - ω 2 * BKLG t - ω 3 * BKLG t-1 ) + ν t+1 (4 ) The joint test of efficiency (i.e., ω 2 = ω 2 *; ω 3 = ω 3 *) is rejected at conventional levels. More important, the results (unreported) indicate that the market prices BKLG t-1 (not BKLG t ) in an inefficient manner. This result is probably not surprising given the high correlation between BKLG t and BKLG t-1. An implication of this finding is that abnormal returns can be earned on BKLG positions for a period longer than one year. Indeed, we do observe (results not reported) the presence of abnormal returns on BKLG dec portfolios for three years into the future, after 23

26 which the abnormal returns are no longer statistically significant. The disappearance of abnormal returns after three years also suggests that an omitted risk factor cannot explain our findings. Recall that the coefficient mapping BKLG to future earnings (EARN), is constrained to be a cross-sectional constant in equation (1). To assess whether our results are sensitive to that assumption, we perform two sensitivity checks. First, we estimate the earnings forecasting equation (1) and the returns equation (4) separately for two main industry groups in our sample durable manufacturers and computers. Separate estimation by industry does not constrain the predictive ability of order backlog for future earnings to be a cross-sectional constant. In unreported results, we find that the market overweights backlog relative to its contribution to future earnings in both industries. Second, we estimate future earnings from BKLG and investigate whether investors misprice such earnings. In particular, we construct a new variable, BKLG-GM, defined as the product of BKLG and the gross margin percentage (1- cost of goods sold/sales) and substitute that variable in place of BKLG in equations (1) and (4). 12 The unreported results of the Mishkin test indicate that the market continues to overweight the relation between BKLG-GM and future earnings. In particular, coefficient ω 2 on BKLG-GM in equation (1) is whereas investors price BKLG-GM at (coefficient ω 2 *) and the difference is statistically significant (likelihood ratio statistic = 9.80, p < 0.00). Taken together, our reported results are not due to the assumption of homogeneity on the conversion of backlog across industries. Finally, we assess whether our results are sensitive to the deflator used to scale order backlog, by using sales as a deflator instead of average total assets in equations (1) and (4). The untabulated results from this procedure are similar to those reported in the paper Do Analysts Appreciate the Implications of Backlog for Future Earnings? 24

27 In this section we explore whether sophisticated market intermediaries, such as equity analysts, understand the implications of order backlog for future earnings when they generate earnings forecasts. Lev and Thiagarajan (1993) identify order backlog as one of the fundamental variables that analysts appear to use beyond GAAP information. Francis, Schipper, and Vincent (2003) isolate order backlog as a leading indicator in the home-building industry based on their assessment of industry reports, analyst reports and popular press articles. Research by Barth, Kasznik, and McNichols (2001) suggests that financial analysts likely aid investors assessment of non-gaap information. Hence, arguments made in the prior literature support the conjecture that analysts incorporate backlog data into their earnings forecasts. On the other hand, other research (Klein, 1990; Abarbanell, 1991; Mendenhall, 1991; Abarbanell and Bernard, 1993; Ali, Klein, and Rosenfeld, 1992) finds that analysts do not fully utilize available information efficiently while setting earnings forecasts. Thus, whether analysts incorporate order backlog in a manner consistent with the cross-sectional association of backlog to future earnings is an open question. To address this question we regress future earnings on scaled decile ranks of backlog and compare the coefficient on backlog ranks with the coefficient from a regression of analyst forecast of future earnings on these backlog ranks. We estimate the following specifications: EPS t+1 = δ 0 + δ 1 EPS t + δ 2 BKLG dec t + µ t+1 (8) FEPS t+1 = λ 0 + λ 1 EPS t + λ 2 BKLG dec t + η t+1 (9) FERR t+1 = (δ 0 - λ 0 ) + (δ 1 - λ 1 ) EPS t + (δ 2 - λ 2 ) BKLG dec t + (µ t+1 - η t+1 ) (10) where FERR is the forecast error computed as the difference between EPS t+1 and FEPS t+1, EPS t+1 is earnings per share for the fiscal year t+1 scaled by stock price at the end of fiscal year t, and FEPS t+1 if the forecasted year t+1 earnings per share made four months after the end of year t 25